wzh19960613/mobileclip2-b-safetensors

mobileclip2-b-safetensors

Chinese documentation: README_CN.md.

Usage is documented in `github.com/wzh19960613/mobileclip2-b-rs`.

This repository contains SafeTensors exports of Apple MobileCLIP2-B weights for a Rust loader, plus the CLIP BPE tokenizer.

Contents

Sources

• —mobileclip2_b.pt: the official Apple full checkpoint from `apple/MobileCLIP2-B`.
• —mobileclip2_b.ts: the Ultralytics text encoder TorchScript export from `ultralytics/assets` GitHub release, used for compatibility with the Ultralytics inference path.
• —tokenizer.min.json: the CLIP BPE tokenizer from `openai/clip-vit-base-patch32`.

Conversion

python convert_ts_to_safetensors.py \
    --input /path/to/mobileclip2_b.pt \
    --component text \
    --output mobileclip2_b.safetensors

python convert_ts_to_safetensors.py \
    --input /path/to/mobileclip2_b.pt \
    --component image \
    --output mobileclip2_b_image.safetensors

python convert_ts_to_safetensors.py \
    --input /path/to/mobileclip2_b.pt \
    --component full \
    --output mobileclip2_b_full.safetensors

python convert_ts_to_safetensors.py \
    --input /path/to/mobileclip2_b.ts \
    --component text \
    --output mobileclip2_b.safetensors

Key layout:

• —text-only: unprefixed text semantic keys, such as token_embedding.weight.
• —image-only: unprefixed image semantic keys, such as cls_token and blocks.0.*.
• —full: text.* and image.* prefixes, plus logit_scale.

License

convert_ts_to_safetensors.py is MIT licensed. The model weights are derived from Apple's official MobileCLIP2-B release and follow the Apple Sample Code License / Apple ML Research Model Terms of Use, for research use only. tokenizer.min.json is sourced from the OpenAI CLIP tokenizer and follows its MIT License.